A research team involving ESILV and the De Vinci Research Center has published a new study in Engineering Applications of Artificial Intelligence.
The article introduces F-XMF, a feature selection framework designed to improve prediction accuracy and interpretability in complex industrial environments.
From industrial data complexity to more reliable predictions
Artificial intelligence is increasingly used to monitor industrial processes and support decision-making. One of the main challenges lies in selecting the most relevant variables from large volumes of data while accounting for relationships between features and limiting redundant information.
A recent publication by Zhiqiang Wang, Associate Professor at ESILV, and his co-authors addresses this issue through a new feature selection framework named F-XMF (adaptive-weighted multi-level feature evidence and redundancy penalty method).
The research has been published in Engineering Applications of Artificial Intelligence, an international peer-reviewed journal published by Elsevier.
A framework combining multiple levels of feature analysis
Traditional feature selection methods often evaluate variables individually, which can limit their effectiveness in complex industrial systems where interactions between variables influence prediction performance.
The F-XMF framework combines several complementary approaches to identify informative features while reducing redundancy. The methodology integrates:
- XGBoost to evaluate the importance of individual variables;
- Attentional Factorisation Machines (AFM) to identify significant interactions between features;
- Conditional Mutual Information (CMI) to assess redundant dependencies among candidate variables.
The framework also incorporates a Differential Redundancy Penalty Mechanism, allowing highly overlapping features to be penalised while preserving variables that contribute useful interaction information.
Adaptive feature fusion for industrial applications
A key aspect of the research is the adaptive weighting strategy used to combine evidence from individual features and their interactions. This validation-based fusion process produces a compact feature subset that supports accurate predictive models while maintaining interpretability.
According to the study, the method improves prediction performance compared with several representative baseline approaches.
The results indicate that combining interaction enhancement with redundancy suppression contributes to more effective feature selection in industrial environments.
Application to the tobacco drying process monitoring
The framework was evaluated using real industrial data from a tobacco drying process. The objective was to predict outlet moisture content, a quality indicator that directly affects process performance.
Experimental results showed consistent reductions in Mean Absolute Error (MAE) across multiple comparisons. The study also demonstrated the framework’s practical value in refining candidate feature subsets and improving prediction reliability.
Beyond this specific application, the methodology can be adapted to a wide range of industrial monitoring and predictive maintenance scenarios in which large datasets contain complex relationships among variables.
An international collaboration between France and China
The publication is the result of a collaboration between researchers from ESILV Engineering School and the De Vinci Research Centre in France, and Wuhan University of Science and Technology in China.
This partnership contributes to ongoing research activities in artificial intelligence, machine learning and industrial data analytics, fields that continue to shape the future of smart manufacturing and intelligent process monitoring.
Learn more about ESILV’s research strategy